Pulmonary impairment independently determines mortality in critically ill patients with acute‐on‐chronic liver failure

In ACLF patients, an adequate risk stratification is essential, especially for liver transplant allocation, since ACLF is associated with high short‐term mortality. The CLIF‐C ACLF score is the best prognostic model to predict outcome in ACLF patients. While lung failure is generally regarded as signum malum in ICU care, this study aims to evaluate and quantify the role of pulmonary impairment on outcome in ACLF patients.


| INTRODUC TI ON
Acute-on-chronic liver failure (ACLF) is a severe complication of cirrhosis and a frequent cause for admission to an intensive care unit (ICU). ACLF is associated with severe systemic inflammation and characterized by acute decompensation of pre-existing cirrhosis, accompanying organ failures and high short-term mortality. 1,2 ACLF patients constitute a heterogeneous group with respect to number and combination of organ failures, aetiology of underlying liver diseases and precipitating events. [1][2][3] The high short-term mortality and heterogeneity of ACLF patients underline the necessity to adequately stratify and identify patients at risk of further deterioration and death. In recent years, the CLIF-C ACLF score has been established as the superior prognostic model to predict short-term mortality in ACLF patients. [4][5][6] The CLIF-C ACLF score ranges from 0 to 100, whereby a threshold above 64 to 70 points is regarded as the futility of care. 4,7 In particular, patients presenting ACLF in combination with mechanical ventilation or pulmonary failure are often considered a vulnerable subgroup in an ICU setting, whose clinical condition can quickly deteriorate, ultimately resulting in death.
These patients are frequently not considered for liver transplantation because of high short-term mortality. 9 To date, however, studies have not discriminated between mechanical ventilation and pulmonary failure with sufficient detail. Also, the exact impact of pulmonary impairment on short-term mortality in ACLF patients remains to be determined. One reason may be that only a small percentage (10-16%) of ACLF patients presented with pulmonary failure. 10,11 The aims of this study are, first to evaluate the role of the need for mechanical ventilation and the presence of pulmonary failure as risk factors for short-term mortality in ACLF patients and secondly to calibrate the CLIF-C ACLF score for pulmonary impairment and thereby improve its predictive performance overall.

| Study design and population
In this retrospective study, 498 patients were evaluated, who were admitted to our ICU/IMC ward between March 2015 and June 2019 with the diagnosis of liver cirrhosis and acute decompensation (AD) or ACLF. Including IMC/ICU readmissions until June 2019, these patients generated a total of 775 admissions. Routinely evaluated clinical data were aggregated for all patients including vital signs, medical history, medication, general clinical data, and laboratory parameters (see Table 1). Data were collected upon admission as well as on days 2, 7, 28, 90 and at 1-year follow-up. If patients left the outpatient care clinic, 1-year follow-up was assessed by telephone interview with the patient and/or the general practitioner to determine survival status.
Patients were identified by using our computerized databases ORBIS (Agfa HealthCare) and MetaVision (iMDsoft). The ethical committee of the University Hospital Frankfurt approved the study (EK 20-707).
Diagnosis of liver cirrhosis was based on clinical data (i.e. laboratory results, sonography and/or other imaging, liver biopsies and endoscopy). ACLF was defined by hepatic and extrahepatic organ failures according to EASL CLIF criteria and ACLF grades were defined according to Moreau et al. 8 CLIF-C ACLF score was calculated according to the EF-CLIF formula. 12 Adjusting the CLIF-C ACLF score for the grade of pulmonary impairment improved the prediction significantly.
Conclusions: This study highlights that not only pulmonary failure but also mechanical ventilation is associated with worse prognosis in ACLF patients. The grade of pulmonary impairment should be considered in the risk assessment in ACLF patients. The new score may be useful in the selection of patients for liver transplantation.

K E Y W O R D S
ACLF, acute-on-chronic liver failure, CLIF-C ACLF score, CLIF-C ACLF-R score, mechanical ventilation, pulmonary failure, respiratory failure Highlights • Mechanical ventilation and pulmonary failure constitute independent risk factors in ACLF.
• The CLIF-C ACLF score underestimates mortality in these patients.
• Adjusting for the presence of MV/PF improves the predictive accuracy of the CLIF-C ACLF score.

Lay Summary
Acute-on-chronic liver failure (ACLF) is one of the main causes of death in liver cirrhosis and is associated with a high short-term mortality. Currently, the CLIF-C ACLF score is the best model to predict mortality in ACLF patients. However, our data suggest that pulmonary complications are not adequately reflected in the CLIF-C ACLF score. This study introduces a calibration variable to adjust for pulmonary impairment. The resulting modified CLIF-C ACLF score seems to have significantly higher accuracy for the prediction of short-term mortality and may help clinical decision-making in the future.

TA B L E 1 (Continued)
Predictors for short-term mortality were evaluated in the overall cohort of cirrhotic patients with first IMC/ICU admission (n = 498) and in an ACLF subcohort of 176 patients, who presented a defined ACLF at first admission (median CLIF-C ACLF 52.7, IQR 45.3-58.3).
In order to evaluate whether mechanical ventilation and pulmonary failure constitute risk factors independent of associated ACLF severity, we performed 1:1:1 propensity score matching. ACLF patients were assigned to one of three groups: receiving high oxygenation  Table 2. Table S1 shows standardized differences (STDs) to express degree of balance of baseline characteristics pre-and post-matching between the respective ACLF subgroups. The matched cohort was derived from ACLF patients at first IMC/ICU admission (n = 176) and was used as a test cohort to compute the calibration variable used for the CLIF-C ACLF-R score. To assess the validity of our results, we assigned patients with ACLF at IMC/ICU re-admission to a separate cohort for internal validation.

| Statistical analysis
Normally distributed data are expressed as mean and standard deviation. Non-parametric data are expressed as median and interquartile range. Non-parametric testing for unpaired comparisons was performed by Mann-Whitney U test for two groups and Kruskal-Wallis-test for >2 groups. Survival rates were analysed using the Kaplan-Meier method. Univariate and multivariate risk factor analyses were performed by Cox regression (backward step-wise likelihood quotient). Predictive performance of prognostic models was evaluated by ROC analysis and Harrel's C-index. The predicted 28day mortality was calculated according to the formula proposed by the EASL-CLIF consortium. 12 A loess curve was constructed to assess the calibration. Standardized mean differences were calculated to assess the balance of covariates before and after propensity score matching. P-values <0.05 were considered statistically significant.

| Introduction of a calibration variable to adjust the CLIF-C ACLF score
We calculated for an easily realizable calibration variable (CV) to adjust for presence and absence of MV or PF and established the CV to be 1 for PF, 0.5 for MV, and − 0.1 for noMV/noPF. The calibration variable was calculated within the matched cohort. The adjusted CLIF-C ACLF score, which we termed the CLIF-C ACLF-R score, is calculated as follows: By factoring in the presence of PF with a CV of an additional 20 points to the CLIF-C-ACLF score, presence of MV with an additional 10 points and noMV/noPF with −2, we achieved a considerably improved predictive accuracy of short-term mortality in the first admission cohort (test cohort) and in the re-admission cohort (internal validation cohort).

| General patient characteristics
The details on the general characteristics of enrolled patients are displayed in Table 1

| Outcome and causes of death
Of our overall cohort of cirrhotic patients admitted to IMC/ICU (n = 498). 20.5% died during the index IMC/ICU admission, with    Tables 1 and 3).
Since the presence of MV and PF are associated with a more advanced stage of ACLF (median CLIF-C ACLF score 51.7 and 55.2 respectively) compared to noMV/noPF (median CLIF-C ACLF 40.9), we performed 1:1:1 propensity score matching, as described in 'Methods'. Balance of baseline covariates pre-and post-matching was assessed by standardized mean differences (STDs, see Sup. Table 1). In the resulting 1:1:1 matched cohort (n = 147, median CLIF-C ACLF 54, IQR 49-59), all subgroups presented comparable age and gender distributions, similar ACLF grade distribution (STD <1%) and comparable CLIF-C ACLF scores (STD ≤10%, see Table 2). Of note, MELD and MELD-Na scores showed markedly higher STD values after matching compared to CLIF-C ACLF scores.
We observed this to be an artificial shift in OF distribution towards 67.3% vs. 38.8%), as shown in Figure 1B and

| Predictive performances of established scores in ACLF patients
To evaluate the predictive performances of different prognostic scores, we performed ROC analysis for 28-day and 90-day mortality. In our ACLF at first admission cohort (n = 176), the CLIF-C ACLF score outperformed all other scores tested for predicting 28-day mortality (AUROC 0.71, 95% CI 0.63-0.79, see Figure 2A). In order to analyse the predictive accuracy of the CLIF-C ACLF score in ACLF patients with pulmonary impairment, ROC analysis was performed separately in ACLF-MV and -PF as well as ACLF-noMV/noPF subgroups (see Figure 2). ROC analysis showed a good prediction of 28-day mortality in all ACLF and ACLF-noMV/noPF patients, with an AUROC of 0.71 and 0.75 respectively. In contrast, in ACLF-MV and ACLF-PF patients, the CLIF-C ACLF score was surprisingly outperformed by the other tested scores. Corresponding 11 C-Indices and 95% CIs are displayed in Sup. Table S7, displaying a comparable trend. An additional ROC analysis was performed to predict 28day and 90-day mortality for each organ failure subgroup, such as liver, kidney, cerebral, coagulation, circulatory and pulmonary failure (see Figure S1). Data showed a fair to good prediction of short-

| Predictive performance of the revised CLIF-C ACLF-R score and validation
Next, we assessed whether the prediction of short-term mortality provided by the CLIF-C ACLF score could be improved, since it fails to adequately reflect mortality in ACLF-PF and ACLF-MV patients.
As described in 'Patients and Methods', we calculated an easily implantable calibration variable derived from a binary logistic regression model to factor in the grade of pulmonary impairment, resulting in our adjusted CLIF-C ACLF R score. In our main cohort of all cirrhotic patients at first admission (n = 498), the new CLIF-C ACLF-R score showed a superior prediction of 28-day mortality compared to the CLIF-C ACLF score in our first admission cohort (AUROC 0.87 vs. 0.81, see Figure 4A,B, Table S4). To illustrate the improved predictive accuracy of the CLIF-C ACLF-R score in ACLF and pre-ACLF patients,  we contrasted the predicted and the observed 28-day mortality of the CLIF-C ACLF and CLIF-C ACLF-R score (see Figure 4). Data shows the CLIF-C ACLF score distinctly underestimating 28-day mortality in the mid-range between 41 and 69 score points ( Figure 4A), whereas the CLIF-C ACLF-R score shows only a minimal overestimation in these patients ( Figure 4B). Supplementary Figure S5 displays a loess-based calibration curve of the CLIF-C ACLF and the modified CLIF-C ACLF-R score, respectively. Shifts in scoring point distribution between the CLIF-C ACLF and the adjusted CLIF-C ACLF-R score, as a result of introducing the calibration variable, are displayed in Figure S2.

| Validation of the CLIF-C ACLF-R score
The calibration variable introduced in this study was calibrated in our 1:1:1 matched ACLF cohort (n = 147), consisting of ACLF patients at first IMC/ICU admission. To validate our findings, we designated the IMC/ICU re-admission cohort as the internal validation cohort to avoid double testing and to reduce confirmation bias. Figure S3 gives an overview of the assigned study groups. In all re-admitted patients (n = 261, AUROC 0.78 vs 0.70) and all re-admitted ACLF patients (n = 148, AUROC 0.74 vs. 0.64), the CLIF-C ACLF-R score showed a significantly higher predictive accuracy for 28-day mortality compared to the CLIF-C ACLF score (see Figure 3A,B respectively). In an external cohort, which was kindly provided by Drolz and colleagues, 21 the CLIF-C ACLF-R score showed a comparable but not superior predictive accuracy for 28-day mortality (see Table S4). In a subgroup analysis, the CLIF-C ACLF-R score showed a better prediction in ACLF grade 3 patients (AUROC 0.62 vs 0.60) and ACLF patients with extrahepatic organ failures (AUROC 0.62 vs 0.60). In addition, multivariate Cox regression in the external ICU cohort showed the new CLIF-C ACLF-R score to be predictive for 28-day mortality, independent of CLIF-C ACLF score, in ACLF grade 3 patients (Exp(B) 1.024, 95% CI

| DISCUSS ION
This study shows that pulmonary impairment, specifically classified according to mechanical ventilation and pulmonary failure, constitutes a distinct risk factor for increased 28-day mortality in ACLF patients, independent of ACLF severity. In our entire ACLF cohort, mechanical ventilation was associated with a 2.45-fold F I G U R E 2 Predictive accuracy of different prognostic scores for 28-day and 90-day mortality depending on presence/abscence of mechanical ventilation and pulmonary failure. ROC analysis performed for CLIF-C ACLF, Child-Pugh, MELD and MELD-Na score for prediction of 28-and 90-day mortality in the overall ACLF cohort (n = 176) and the ACLF-MV (AUROC 0.67/0.66), ACLF-PF (AUROC 0.49/0.52) and ACLFnoMV/noPF (AUROC 0.74/0.67) subgroups. (or SpO2/FiO2 < 214), according to the CANONIC study design. 16,17 In several studies, mechanical ventilation has been identified as a risk factor in cirrhotic and ACLF patients, associated with increased mortality. 11,18 As can be expected, the duration of mechanical ventilation and the time point of initiation are predictive of patient outcome, whereasearly mechanical ventilation discontinuation seems to be associated with better outcome. 11,16 However, the subgroup analyses in previous ACLF studies could neither dissect nor quantify the role of presence or severity of pulmonary failure as an independent risk factor of mechanical ventilation. [11][12][13][14][15][16][17][18] To specifically analyse the role of mechanical ventilation and pulmonary failure, we adjusted for sex and severity of liver disease to eliminate possible confounders. Importantly, in this propensitymatched cohort, pulmonary failure almost doubled 28-day mortality compared to patients without pulmonary impairment. Patients with mechanical ventilation showed a more than 50% increase in 28-day mortality. These findings identify pulmonary impairment as a critical clinical marker for poor prognosis independent of ACLF severity.

F I G U R E 3 ROC analysis of the CLIF-C ACLF and adjusted
The CLIF-C ACLF score was introduced in 2014 and it was derived from the CANONIC study. 12 It has since been established as the superior prognostic model to predict short-term mortality in ACLF patients. 3,6,19 Overall, we were able to confirm that the CLIF-C ACLF score is the best prognostic model for ACLF patients. However, the lack of precision regarding pulmonary impairment may render this score less suitable for critically ill patients requiring mechanical ventilation. 9 Thus, suboptimal risk stratification could decisively F I G U R E 4 Observed and predicted 28-day mortality of the CLIF-C ACLF and CLIF-C ACLF-R scores. (A) Histogram of the predicted and observed 28-day mortality of the CLIF-C ACLF score. (B) Histogram of the predicted and observed 28-day mortality of the CLIF-C ACLF-R score.
affect ICU decision-making, that is regarding transplant allocation. In our study, the CLIF-C ACLF score showed poor prediction of mortality for ACLF patients with pulmonary failure, similar to the flipping of a coin. In ACLF patients with mechanical ventilation but without defined pulmonary failure, the CLIF-C ACLF score performed better but was still outperformed by MELD, MELD-Na and Child-Pugh score. In order to improve the predictive accuracy for this specific patient population, we introduced a calibration variable adjusting for pulmonary impairment. This revised CLIF-C ACLF-R score improved overall predictive accuracy for mortality.
Notably, the investigations of the CANONIC study also demonstrated that the CLIF-C ACLF score underestimated mortality in patients with a score ranging up to 64 points. 12 We confirmed this in our dataset. However, the discrepancy was reduced after introducing the calibration variable. At least for our cohort, the underestimation of the CLIF-C ACLF score seems to be because of the lack of precision for mechanical ventilation or pulmonary failure.
To internally validate our results, we designated an ACLF cohort separate from the initial test and calibration cohort, which consisted of all IMC/ICU re-admitted patients. In this validation cohort, we were able to internally confirm the increased predictive accuracy of the new CLIF-C ACLF-R score. In addition, we validated our results externally in a large ICU cohort. 21 While the CLIF-C ACLF-R score showed comparable prediction for short-term mortality in the overall external cohort, subgroup analysis indicated that the prognosis of patients with a higher number of organ failures may be improved. However, sampling variability in our goodness-of-fit analysis shows that further evaluation and validation of the proposed CLIF-C ACLF-R score in other clinical cohorts is necessary.
This study has several limitations. First, as a result of the retrospective design results from this study could be affected by an inherent inadequacy of collected data. Nonetheless, we are confident that the data in this large retrospective cohort can be viewed as robust, since mortality rates, organ failure rates and complication rates are consistent with available data, that is as published in the CANONIC and PREDICT study. 3,13,17,20 By propensity score matching and adjusting for ACLF severity we aimed to effectively reduce selection bias in our cohort. Although data showed successful balancing for the selected confounding covariates in our propensity matched cohort, namely the CLIF-C ACLF score and sex, we acknowledge that balance of all baseline covariates (>30) was not fully achieved. Depending on the clinical data, reaching full balance among all baseline covariates can be difficult to obtain without substantially sacrificing sample size. Thus, we acknowledge that balancing can be seen as an inherent limitation to interpretation.
Importantly, however, we found the MELD score STDs between our matched subgroups to remain >10% as an artificial effect of patient grouping, as was described earlier. Instead of adequately reflecting differences in disease severity between subgroups, to the authors this observation rather highlights the limited capability of the MELD/MELD-Na score to fully capture predicting factors of mortality in ACLF patients. This is underscored by the fact, that alignment of CLIF-C ACLF score and MELD score STDs in full balance rendered unfeasible in our cohort, when excluding a non-MELD-captured OF by patient subgrouping, regardless of PS condition. However, if this would be considered as a confounder, differences in MELD scores between matched subgroups would rather result in an underestimation of the observed effects on mortality. In regard to calibration, results of internal validation and the external ICU cohort should be considered with caution. To eliminate multiple testing and avoid confirmation bias, we designated a separate cohort for internal validation. However, an inherent bias in our internal validation cohort cannot be ruled out in a retrospective study of this nature.
Furthermore, general characteristics of the patients included in our study differ from those of the external ICU cohort but are similar to the CANONIC study. This might contribute to the reason as to why we can validate our findings only in the most severe ACLF patients.

| CON CLUS ION
In summary, we were able to identify and quantify the role of

ACK N OWLED G M ENT
Open Access funding enabled and organized by Projekt DEAL.

CO N FLI C T O F I NTE R E S T S
The authors have no conflicts of interest to disclose.

WR ITI N G A SS I S TA N CE
The authors have no writing assistance to disclose.

PER M I SS I O N S TO R EPRO D U CE M ATER I A L
The presented data in this manuscript is derived from original data.